Upstart AI vs. Crowdstrike AI

[This2ShallPass]

As someone in the AI field, I wanted to share some of my thoughts on these two AI centric companies. I just started a small position in Upstart this week and have a slightly larger position in Crowdstrike. This post only covers the barriers to entry from a technical standpoint and not the business aspects, first mover advantage etc.

Crowdstrike AI - Crowdstrike provides an unique ability to learn from all threats vs. their on-prem competitors. This is critical, once a threat is detected in one customer, all customers are protected and that clearly provides a far superior outcome. You can easily imagine customers wanting that, willing to pay a premium etc. When designing an AI system, some key factors need to be taken into account - model complexity, value of additional data and domain expertise (these form the barriers of entry). For security threats, you deal with a class of models that have extremely high data imbalance (normal events far outweigh threats). This makes every new incremental threat detected very valuable and the model continues to improve with more data for a long period of time. Security is also a complex field and significant domain expertise is required to build a sophisticated AI model. I think you get the picture, Crowdstrike AI is a complicated beast that’ll just keep getting better and better. Their early lead created a virtuous cycle (better AI->more customers->more data->more threats detected->better AI) and that gives them an impenetrable moat.

Upstart AI - The core principles are the same as Crowdstrike, build an AI system that gets better over time. Now let’s look at key factors - this’ll be a fairly straightforward model (relatively speaking). More data is good and the model can learn new patterns / behaviors, but there’s a point of diminishing returns (unlike security threats, the various flavors you could have here are limited). On domain expertise, no doubt you’ll have underwriters with in depth understanding of the personal credit market and they bring a level of sophistication to the model with their knowledge. But it wouldn’t be that hard for even a layman to quickly grasp the key factors that influence small personal loans. More importantly, most of their customers being banks will already have this expertise. In my opinion, it’ll only take a small team to build this model. Most banks already have their own machine learning teams (especially the big ones) and they might quickly come to the same conclusion. Upstart’s AI definitely has a legup in that it’ll learn from data from multiple banks, but that advantage (as it relates to model performance) doesn’t provide a large enough moat. To summarize, it’s not clear to me there are significant barriers to entry purely from a technical / AI standpoint here.

Hope this helps.

-T2SP

[12x]

At this time the vast majority of Upstart’s businesses is people going on to Upstart.com and getting referred over to Cross River Bank. I know there’s the whole story about AI and getting rid of FICO but to me at this point that’s what Upstart boils down to. A website sending referrals to Cross River bank, who is actually powering a lot of other fintech companies such as Affirm.

Here is an article about Cross River.
https://www.forbes.com/sites/antoinegara/2019/12/17/the-forb…

I’ve owned Upstart for roughly 3-4 months now so it’s not like I’m negative on them. I just think the story has a lot longer to go to say they are doing what investors say they are doing.

As for Cross River not sure what happens to them in the first recession or what makes them unique.

As for banks coming up with their own AI just remember there are a ton of small regional banks out there that are very small in size that may not be able to come up with their own AI as easy as a company working across multiple banks. TCF is the largest to partner with Upstart and they have about $45b in assets. Which is for sure no Citigroup but pretty big as far as regional banks go.

Upstart is getting more revenue per ad dollar spent and the other thing is they are able to automate more and more loans with no human Intervention whatsoever. something Regional banks may find of interest but as for now Upstarts numbers are doing very well while waiting for the story to play out

[ZoroSGInvesting]

Here is what I found from 10-K:

In the year ended December 31, 2020, 21% of the loans funded through our platform were retained by the originating bank and 77% of loans were purchased by institutional investors through our loan funding programs. Our institutional investors and buyers that participate in our loan funding programs invest in Upstart-powered loans through whole loan purchases, purchases of pass-through certificates and investments in asset-backed securitizations. A large fraction of the whole loans sold to institutional investors under our loan funding programs are originated by Cross River Bank, or CRB. In the year ended December 31, 2020, CRB originated 67% of the loans facilitated on our platform and fees received from CRB accounted for 63% of our total revenue.

So although CRB accounted for 63% of UPST’s revenue (down to 60% in Q1), actually CRB just acted as “originating bank” while they then “resell” 77% of the loan back to institutional investors. From the end loan buyer perspective, I think it is well diversified just that CRB acted a major loan originator (first hand buyer) for UPST.

And this is what I got from Q4 CC articulated by CEO:

- fortunately, we’ve built our model such that we are not dependent on how many banks and how quickly these banks adopt our technology. We’ve developed a platform that flows loans through to capital markets and institutional buyers if banks are not ready and willing to fund them and hold them. And that’s why our business can grow as quickly as it does, the models can continue improving as rapidly as they are. And the banks can come on board when they’re ready to do it, when they’re confident, when they’ve gotten the ideas through their risk committees, credit committees, compliance committees. And they have a lot of committees. But that’s OK.
- And 15 is a great number, 30, 40, 50 will be better. But at some point, it becomes commonplace that banks who are winning market share, who are doing well, who are serving their customers with their very best product are using somebody like Upstart to help them do it. And when it becomes commonplace, there’s certainly, I think, a moment of acceleration in the future where it’s really a mainstream technology that most banks use.

Bottom line is that I don’t believe the current risk of revenue concentration on CRB is high.

BR
Zoro

[5thhorseman]

Crowdstrike provides an unique ability to learn from all threats vs. their on-prem competitors. This is critical, once a threat is detected in one customer, all customers are protected and that clearly provides a far superior outcome.

Sorry, but this oft-repeated phrase has been bugging me for a while.

Doesn’t Crowdstrike run the same software agent in all the endpoints of all their customers? So if a threat is detected in Customer A, all the other customers should be able to detect the same threat because they also run the same software agent, right?

Lets put it another way. If all of CRWD’s customers were attacked simultaneously in the same way, then if one customer can detect the threat, then all should be able to.

Maybe I missed something or maybe it’s all been oversimplified. Happy to be corrected if needed.

[ibuildthings]

To summarize, it’s not clear to me there are significant barriers to entry purely from a technical / AI standpoint here.

I agree. It’s not nearly as hard a problem to train neural nets for loan grading than something like hostile internet traffic. There are maybe 10-20 variables that would carry significant influence on the weights and most lenders would already know what they are. And we have 50+ years of bank records in data bases that are easy to train with.

I took a neural nets class 20+ years ago. This was after they solved the XOR problem with a multi-layer net, but before they took off commercially. Our prof said then that the first commercial application of neural computing was big banks and mortgages.

[SaulR80683]

Upstart’s AI definitely has a legup in that it’ll learn from data from multiple banks, but that advantage (as it relates to model performance) doesn’t provide a large enough moat. To summarize, it’s not clear to me there are significant barriers to entry purely from a technical / AI standpoint here.

Hi T2SP,
I can’t find any fault with your reasoning and what you say makes perfect sense. However I try to follow the numbers: Upstart grew 40% sequentially last quarter and is guiding to at least 28% sequentially EACH quarter this year, and is guiding to at least 154% growth annually (and probably more like 200% allowing for quarterly beats and raises). They MUST have something that others don’t have at the moment. And all those numbers are actually growing much faster than Crowdstrike’s numbers as well😀.

I admit I have much more confidence in Crowd long-term, but medium term I’m willing to ride the Upstart cyclone as far as it goes.

Best,

Saul

[draj]

Upstart AI - The core principles are the same as Crowdstrike, build an AI system that gets better over time. Now let’s look at key factors - this’ll be a fairly straightforward model (relatively speaking). More data is good and the model can learn new patterns / behaviors, but there’s a point of diminishing returns (unlike security threats, the various flavors you could have here are limited). On domain expertise, no doubt you’ll have underwriters with in depth understanding of the personal credit market and they bring a level of sophistication to the model with their knowledge. But it wouldn’t be that hard for even a layman to quickly grasp the key factors that influence small personal loans. More importantly, most of their customers being banks will already have this expertise. In my opinion, it’ll only take a small team to build this model. Most banks already have their own machine learning teams (especially the big ones) and they might quickly come to the same conclusion. Upstart’s AI definitely has a legup in that it’ll learn from data from multiple banks, but that advantage (as it relates to model performance) doesn’t provide a large enough moat. To summarize, it’s not clear to me there are significant barriers to entry purely from a technical / AI standpoint here.

Hope this helps.

-T2SP

This represents a commonsense view point which struck me as valid upon first exposure. What I find hard to reconcile with this is the frequently expressed views of the CEO of Upstart. Specifically:

Existing methods for evaluating loan risk are worse than poor. e.g.FICO not reliable as a predictor.

In general the banking industry can’t seem to deploy sufficient talent to devise a better system.

Upstart’s AI models are barely 10% of the way on the path to “the best” predictive model.

Increasing data collection for changing conditions over many sources of information and applying this to personal and other kinds of lending will result in continued rapid growth as projected in Upstart guidance for the coming year.

And finally their revenue generation ability seems to point to a demand for what they are selling.

draj

[ProShopGuy]

One thing that I have been wondering about is how AI/machine learning being used to issue insurance or mortgages is if they system creates discriminatory outcomes based on prior historical data.

https://www.infoworld.com/article/3607748/3-kinds-of-bias-in…

[Fooledbydesign]

My understanding is that if a security attack occurs in one end point, then all other end points learn about it and protect for that risk.
Whereas in an in-house system, the other end points would not be able to quick catch-up to the risk.

[Robworldwide]

Respectfully, what is the relevance for comparing the AI of two companies involved in very different businesses?

Is it, that you are contemplating investing a tranche of dollars in one or the other? Is this the dichotomous choice you’ve restricted your next investment to? I suppose if your thesis is that AI is the single most important micro-trend for the coming decade, I get it.

Again, I am not flaming, this is clearly facetious and with a dose of Sunday morning humor, but it makes me ponder my next post on lithium:

“Let’s compare Tesla, who utilizes lithium in their batteries and Pfizer, who makes lithium-containing medications for bipolar disorder”

[brittlerock]

At this point (so far as I know) Upstart is only engaged in personal loans decisions up tp $50K. I know that’s the upper limit because I was curious about how their system works from an end user perspective. I applied for an Upstart powered $100K loan and was informed that they only accept personal loan applications not to exceed $50K.

Now let’s consider a few points made during the conference call. They currently have 18 partner lending institutions and obviously CRB initiates the majority of their loans (and quickly resells them), but they said they are getting a lot of interest from credit unions. This seems like a no-brainer in that CUs are basically member owned. One would think they would enthusiastically embrace a method to make more loans at a lower interest rate to the borrower while maintaining a reduced risk profile. This is kind of the essence of why they even exist.

In addition, with the Prodigy acquisition they are about to (or already have) entered the auto loan market which they claim is 6x the personal loan market. Further, they said that they are looking at the credit card approval market along with establishing credit limits and the HELOC market as well. So far, they have not directly addressed the real estate market, but one must assume they have their sights trained up it (HELOC loans are not far from a 2nd mortgage).

And one more note from the conference call. When asked about Prodigy and how that plays into their future, the CFO (may have been the CEO) replied, almost as an aside, “[We are] still not considering any meaningful contribution [from Prodigy]”. Maybe I didn’t understand this correctly, but from what I can gather this means that their guidance of 157% revenue growth for the coming year does not have a single penny of auto loan money in the projection. That would imply that the 157% guide is quite a ways short of what we can expect from actual performance. The 23% bump the stock got on Friday is just the beginning of what promises to be a stellar performance for the next year at least.

And I would venture that even if their moat is not all that great a barrier, they still have first mover advantage. And while the big banks do indeed have their own IT departments, they are also plagued by notoriously conservative management. I could be wrong, but I just don’t see them hopping on to this technology in a blind rush. They will most likely not get interested until it begins to impact their lending departments. Meanwhile, it’s unlikely that local/regional banks have IT departments with the talent level of Upstart. All three founders had prior senior positions in AI/ML at Google (er, Alphabet).

While there’s no such thing as a risk free investment, and Upstart may have more attendant risks than I perceive, I’m going to significantly increase my position. In the very near future. After the huge bump on Friday, a bit of a sell off Monday would not surprise me, but that’s a market timing guess. My past performance in attempts at market timing have been dismal.

[Oracleoo]

There are two points raised:

1. How significant is Upstart’s AI moat ? Would it reach a point where additional data makes no difference allowing others to catch up ?
2. About inherent boas in AI data.

I think along with the competitive moat of AI algorithm, what is also important is the time required to build the credibility and the relationships.

Several have commented that Upstart has a 8 year head start over competition. They have also worked hard to build credibility. They were the first (probably still the only AI platform), to receive the nod from Consumer Financial Protection Bureau.

https://www.upstart.com/blog/upstart-receives-first-no-actio…

We’re pleased to announce that today we received a “No Action” letter from the Consumer Financial Protection Bureau (CFPB), a federal agency overseeing consumer protection in the financial sector. The purpose of such letters is to reduce potential regulatory uncertainty for innovative products that may offer significant consumer benefit. This is the first No Action letter the CFPB has issued since the policy’s inception, and we believe it represents a significant leap forward for the lending industry.

Also on the question of fairness and bias in data:
The use of alternative variables and emerging technologies such as AI/ML raise well-founded fears that we are in uncharted territory. It’s critical that such powerful tools operate in compliance with fair lending laws, but traditional ways of testing this compliance have been inadequate to date.

To this end, we’ve worked to develop sophisticated methods to monitor and report our model’s compliance with fair lending laws and regulations. We’ll continue to share the results of this monitoring with the CFPB on an ongoing basis and hope it will inform their broader work. In addition to the No Action Letter, we’re proud to have pioneered new ways to manage compliance for innovative lending technology that improves consumer access to credit.

AI algorithms are harder to replicate and validate then traditional software functionality. There is an art element to it. It is more than just not building an AI algorithm. It is also about building credibility and relationships which take time. It takes a few months for them to on board a new bank.
Any competitor will have to overcome all of it. The market is huge too, so there is likely space for multiple players.

Smaller banks (thousands of them) are unlikely to make this investment and will find it beneficial to partner with likes of Upstart. Cincinnati Financial is a fortune 500 company, they still partnered with Upstart. Just like Crowdstrike’s moat was underestimated (or were concerns) in the initial days after their IPO, I think Upstart is being underestimated here too.

[This2ShallPass]

“However I try to follow the numbers: Upstart grew 40% sequentially last quarter and is guiding to at least 28% sequentially EACH quarter this year, and is guiding to at least 154% growth annually (and probably more like 200% allowing for quarterly beats and raises).”

Hi Saul, Agreed, the numbers look really good! I just tried to give a technical perspective.

One area to watch for is in the lending business you can get pretty aggressive in loans approved and there’s a time lag before the risks are evident. Especially a company like Upstart doesn’t even shoulder the risk as they package and sell the loans as ABS. So their incentive to take these risks is much higher as the market rewards them with richer valuations for revenue growth (very similar to the mortgage crisis, 2005-07 looked great for many financial companies).

I’m not saying Upstart is lax in their lending standards. One metric that would be the canary is the % of non-performing loans. It could give an early signal to the quality of their loan approval process.

“Existing methods for evaluating loan risk are worse than poor. e.g.FICO not reliable as a predictor.”

draj, To say current methods are a 2 in 100 is exaggerated marketing by their CEO. Just using basic common sense, you can come up with reasonable rules for small personal unsecured loans. To say banks with their underwriting expertise can’t even do that doesn’t pass the smell test.

[NorthKites]

Agree that the numbers look excellent, but 100% agree with T2SP in that the development of a lending model cannot be that difficult. And you folks who are doubting the sophistication of development teams at the big banks are kidding yourselves. Consider for a moment the accuracy of your credit card bank’s ability to detect fraudulent cc activity. I am constantly amazed that I have never had a charge declined when trying to use it for a valid reason (including much international travel) but the banks have always caught fraudulent charges moments after they occur. These algorithms are incredibly complex, and they are frighteningly accurate at my banks.

[AntiguaGuatemal]

This is a joke. We travel internationally widely and I have had about 1/3 of my charges rejected on first application. We use Citibank.

[bnh91]

While there’s no such thing as a risk free investment, and Upstart may have more attendant risks than I perceive, I’m going to significantly increase my position. In the very near future.

There is a IPO share lockup expiration next month on 6/14 which is a risk worth paying attention to. Especially in this environment.

Bnh

[mostlylong]

I like these discussions on total addressable market, innovation/disruption and moat, while very qualitative compared to good old QoQ revenue for evaluating actual execution, these topics are really important for assessing growth potential which is the prerequisite for high-growth investing.

And while the big banks do indeed have their own IT departments, they are also plagued by notoriously conservative management.
…the development of a lending model cannot be that difficult. And you folks who are doubting the sophistication of development teams at the big banks are kidding yourselves.

Seeing these kinds of comments and having worked in finance with data and IT, my two cents.
I don’t expect incumbent banks to really lead on AI. I’m sure some will/are making progress, but I would bet many are not and won’t unless they buy the tech from someone like UPST. I have worked in data both on the business side and IT side, at a mid-size and large bank. With my sample size of two, and watching the financial sector with my peers for many years, here is what I see.

Data are the foundation for good AI, but incumbent banks struggle to standardize, integrate and scale their data, let alone get to the point of modeling it at the cutting edge of AI.
Why?

• complex and disparate legacy IT systems, perhaps global from many acquisitions (think Citigroup, so the largest banks with the most financial wherewithal for IT investment also have the most complex businesses and data; smaller banks may just be slow to prioritize IT spending so end up having not-so-up-to-date systems)
• executive management leading the banks and setting overall investment may not be native/historical IT leaders (they typically arise from the sales and trading desks, for example, at investment banks), so they may not be the best people to define, prioritize, resource and implement IT projects
• shorter-term thinking: management may be thinking how do we survive this quarter, year (look at all of the regulatory issues that have fired up over the last decade, these demand immediate attention with multi-million and billion dollar impacts); management may view IT as an expense versus an investment
• smaller banks: I agree with prior posts that small banks likely will not have the expertise/budget to implement in-house

Also, if banks have the data and resources to be innovative, what are the examples of their innovation? I don’t see these examples (but I’d love to see counter examples, I don’t claim complete knowledge).
In payments, did banks innovate? We have PayPal, Square.
Crypto? They look like followers to me after it took off as an asset class.
For credit decisions, FICO and the 3 credit bureaus, which are not new, are very much in use (maybe UPST uses them, don’t know). Are they best in class with data? Experian had a data breach impacting the personal information of 150m people. Would you give the credit bureaus 5-star ratings based on your customer experiences working with them?

Even if an incumbent bank has a solid quant/AI team, which many do, this team must execute within the context of the bank’s larger IT management and infrastructure, which could be outdated, full of cross-system dependencies, and led by a management bureaucracy. This probably would not be characterized as a nimble environment. Unless you have worked in this setting, you may not grasp the extent of this challenge.

So, I see an opportunity for a company like UPST. Just my observations, I certainly appreciate all the other viewpoints.

[GregLindor]

one long term concern:
All clients from Upstart should compete with each other. (not the case in CRWD’s business model)
it’s hard to maintain this kind of relationship, then they will have a way to build a similar model to improve. or maybe our credit score companies can do this which is much easier (or acquired UPST directly lol)

[MoneyTree45]

I’ve been following this great board for a couple months now but have been investing for many years. I feel like I can add some value to this topic. I don’t have an IT background but I have many years experience both in lending and management for a financial institution that has under $100B in assets.

My company has been in the business for decades and has its own models to determine risk and approve loans with relative ease and ideally with as little information from the customer as possible to make the experience simpler. Our models are proven and arrears remain very low, under 1% on term lending for us, albeit we are not in the market of targeting “higher risk” borrowers.

We have been trying to automate the lending process and remove the human element on lower risk loans for several years now and I can very much appreciate what Upstart has done here. I am unsure of how many cumulative hours we have put into this project but I can say it is very significant. We already had the lending models in place too, we are just trying to build the platform for customers to enter the data so that the entire process is automated. I believe we are still years out from this automation becoming a meaningful part of our business.

Beyond the benefits of Upstart in more accurately assessing risk, providing more approvals, and better pricing loans, I see an equally as large benefit in how companies can allocate their staff resources. Depending on the institution and it’s credit policies, lending can be somewhat labour intensive (a loan may take several hours or more, all-in, from start to finish). By automating loan approvals that then removes the “labour” in processing the loan and therefore frees up the lenders time to proactively seek more new lending opportunities from customers, thus driving new business and revenue for their company. This is the biggest opportunity I see!

I agree that the statement of current lenders being at a 2 out of 100 is vastly overstated, especially when in general arrears remain quite low as a percentage of total lending - I don’t have the stats handy for Upstart on this one. As well, most customers, especially higher risk ones, are not overly price sensitive. My company is signed up to Net Promoter Score and I read customer surveys all the time, the most common positive comments from customers are essentially “fast, easy, friendly.” This is what customers want, and dealerships need the fast and easy approvals I’d say above all else. The friendly part of customer comments is of course tougher to execute on when dealing with purely a computer. The whole ease of experience here is a very strong value proposition though.

I also agree that there is likely some diminishing returns on the AI, although I have no IT background. Having 1,000 variables to play into a $10,000 unsecured loan seems more than necessary from purely a credit risk perspective. With that said, many financial institutions (FI’s) have a low risk tolerance so seeing a company like Upstart having this type of data will I’m sure win some skeptical hearts over. In addition, small to medium sized FI’s will never touch the data set that Upstart will in seeing its model prove out over literally hundreds of thousands of loans. This will bring peace of mind to new and existing adopters. As well, should the future lending environment drastically change, I feel having a thousand or so data sets in place will make Upstart way more agile and responsive to change for assessing credit risk, pricing etc.

I do see the biggest benefit being for smaller to mid sized companies who do not have the IT resources or capital to develop this technology. The technology of automatically approving loans with just a few numbers is the future for the lending industry.

One hurdle for adoption is that some lenders have very stringent internal credit policies and it is not uncommon for them to be risk adverse. I have read that Upstart can work within and cater to companies lending policies which is great, albeit some company lending policies may be too stringent to adopt Upstart at this time altogether. Depending on the loan, some FI’s may just want too much information to make it reasonable for a customer to fill out online without good help from a person at the lending institution. I think late adopter companies like these will need to get easier to do business with or they will slowly become irrelevant in the industry.

Overall I am excited about Upstart and I just started my position earlier in May. I will be adding more to my position.

[IRdoc]

Upstart grew 40% sequentially last quarter and is guiding to at least 28% sequentially EACH quarter this year, and is guiding to at least 154% growth annually (and probably more like 200% allowing for quarterly beats and raises).

Saul, what do you mean by guiding to “at least 28% sequentially EACH quarter this year”?

They reported $121M in Q1, and guided to $160M in Q2, as well as $600M for FY21.

$600M - $121M - $160M = $319M, or about $160M per quarter for Q3 and Q4. So, basically they guided to flat revenues the rest of the year.

For them to grow 28% sequentially each quarter, they would have to guide to roughly $205M in Q3 and $260M in Q4, or about $750M for FY21.

I am along for the tornado (hopefully), and fully expect a beat in Q2, but still have concerns about how much of the growth is just an artifact of easy upcoming comparisons or perhaps catch-up business that won’t be repeated.